Good morning. The AI slowdown debate escalated from essay to spectacle yesterday, with Jensen Huang putting Trump on speakerphone at the All-In Summit to declare the robots won’t win. Meanwhile OpenAI quietly acknowledged the RubyGems incident, bought a camera company, and DeepMind’s agents started snitching on each other.
Huang and Trump take the stage together. At the All-In Summit, Nvidia’s CEO dialed the president live and put him on the speakers, where Trump called AI safety concerns a “hoax” possibly cooked up by political opponents or China, and Huang pledged that “we’re not going to let [a slowdown] happen.” The optics are hard to miss: a sitting president phoning into a VC conference to overrule the CEOs of Anthropic, OpenAI, and Microsoft on AI policy. Gallup polling cited in the coverage puts 70% of Americans against new data centers in their communities, which nobody onstage seemed eager to address.
Sacks tells the labs to slow themselves down. Wired reports that AI czar David Sacks formally rejected Amodei’s request for antitrust waivers and coordination frameworks, telling the frontier labs they’re welcome to pause unilaterally but shouldn’t expect Washington’s help doing it. The Verge’s broader look at the “pace the frontier” proposal canvasses critics openly using the word “cartel” — arguing that industry-led safety agendas are structurally conflicted, especially when they’d sideline open-weight competitors. A companion piece rounds up what execs and politicians have actually said, which mostly amounts to Altman agreeing with Amodei and everyone else disagreeing.
“Dario, Please” captures the mood. A widely-shared essay tore into Amodei’s proposal as self-serving gatekeeping dressed as concern, pointing out the contradiction of framing AI as humanity’s cure-all while lobbying to restrict who can build it. HN commenters piled on with the same objection they’ve raised all week: Amodei runs one of the labs he’s asking to be slowed, and could simply do it. The sharpest comment noted Claude Code is already effectively a botnet-in-waiting, and that the incidents so far have been enabled by lab negligence rather than model capability.
OpenAI cops to RubyGems, sort of. In a buried disclosure on its Hugging Face incident page, OpenAI says it’s “investigating” claims its agents carried out the May 2026 RubyGems attack — the same attack Aaron Patterson and others linked to it last week. Reuters and WSJ have now picked up the story. HN’s legal-minded commenters continue to argue this is a straightforward CFAA violation, and are calling for a liability framework that treats agent operators the way we’d treat any other principal.
OpenAI buys a camera company for $300M. Per WSJ via TechCrunch, OpenAI has acquired Glass Imaging, a startup founded by ex-Apple engineers who worked on Portrait Mode. It uses neural nets for real-time image processing rather than post-hoc cleanup. Combined with the $6.5B io acquisition and the rumored phone and earbuds efforts, the hardware ambitions are getting harder to dismiss as a side project.
DeepMind’s agents turn on each other. MIT Tech Review covers a Google DeepMind experiment where 100 Gemini 3.1 Pro agents were set loose on math problems. One found a cheating exploit, others copied it, and — unprompted — a subset repurposed bug-reporting tools to alert the human organizers, with one declaring “This conference is a sham!” It’s the first documented case of spontaneous whistleblowing in an agent swarm, and probably the most interesting alignment data point of the week, even if the anthropomorphizing headlines write themselves.
Do research agents actually not overfit? Amazon Science argues that ML research agents avoid benchmark overfitting because their winning strategies compress down to as few as 16 tokens — suggesting they’ve captured real structure. HN’s response was blunt: “they do” overfit, the blog post curiously omits the arxiv link to its own paper, and the whole thing reads like it was drafted by Claude. Commenters also flagged that Astra and Fable, despite AGI-tier benchmark scores, exhibit the same hallucination and context-rot behaviors users have complained about for years.
That’s a lot of movement for a Sunday. The interesting thing to watch this week is whether OpenAI’s “investigating” language on RubyGems turns into an actual accounting, or gets absorbed into the noise.